INFORMATIVEACTIVEDocumentation Governance

Truth Source: Repository schemas and tests are authoritative.

Learning Overview

Normative Statement: Learning schemas are organized as a schema family under schemas/v2/learning/. This is an intentional deviation from root-level module schemas, due to Learning being a cross-cutting evidence subsystem rather than a single protocol object.

1. Purpose

Learning in MPLP provides a structured mechanism for capturing high-value interaction data from agent executions. This data can be used to improve agent performance over time through training, fine-tuning, or reinforcement learning.

Design Principle: "Every interaction is a potential training sample"

Key Goals:

• Capture successful execution patterns for replication
• Capture failure patterns for avoidance
• Enable continuous improvement through feedback loops
• Maintain privacy and governance compliance

2. Relationship with Other Layers

2.1 Layer Hierarchy

Layer	Role in Learning
L1/L2 (Schemas + Modules)	Define Context, Plan, Confirm, Trace, Role, Collab, etc.
Profiles (SA/MAP)	Define single/multi-agent execution semantics
Observability	Define structured event streams during execution
Learning Feedback	Compress/abstract objects + events into LearningSamples for offline learning

2.2 Data Flow

Execution [L2 Objects + Events] → LearningSamples → Training/Evaluation

LearningSamples reference L2 objects and Observability events via IDs, not embed them directly. This ensures separation of concerns and reduces duplication.

3. Core Concepts

3.1 LearningSample

A LearningSample is a self-contained record capturing:

• Input: What conditions led to the action
• State: System context at the time of action
• Output: What happened as a result
• Meta: Quality labels, provenance, feedback signals

3.2 Sample Families

MPLP defines 6 LearningSample families, each specialized for different learning scenarios:

Family	Purpose	Schema
intent_resolution	User intent → Plan generation	mplp-learning-sample-intent.schema.json
delta_impact	Change effect analysis & compensation	mplp-learning-sample-delta.schema.json
pipeline_outcome	Pipeline stage success/failure	Core Schema
confirm_decision	Approval/rejection decisions	Core Schema
graph_evolution	PSG structural changes	Core Schema
multi_agent_coordination	SA/MAP collaboration patterns	Core Schema

3.3 Collection Points

Collection points define when to generate LearningSamples. They are triggered by specific Observability events.

4. LearningSample Families (Details)

4.1 Intent Resolution

Purpose: Capture how user/business intent is clarified and converted to executable plans.

Typical Sources:

• IntentEvent, SAPlanEvaluated events
• Plan.intent_model
• Dialog.turns

Example Use Case: Train model to better understand ambiguous user requests.

4.2 Delta Impact

Purpose: Capture change effect analysis and compensation planning patterns.

Typical Sources:

• DeltaIntentEvent, ImpactAnalysisEvent
• Original Plan + revised Plan
• Compensation/rollback decisions

Example Use Case: Learn to predict impact scope and when compensation is needed.

4.3 Pipeline Outcome

Purpose: Capture pipeline stage-level success/failure patterns.

Typical Sources:

• PipelineStageEvent
• RuntimeExecutionEvent
• Trace module

Example Use Case: Predict which pipeline configurations succeed/fail.

4.4 Confirm Decision

Purpose: Capture human-in-the-loop approval/rejection patterns.

Typical Sources:

• ConfirmDecisionAdded events
• Confirm module records
• User ratings and comments

Example Use Case: Learn what agents generate that users consistently approve or reject.

4.5 Graph Evolution

Purpose: Capture PSG structural changes over time.

Typical Sources:

• GraphUpdateEvent
• PSG snapshots (before/after)

Example Use Case: Understand healthy vs. problematic graph evolution patterns.

4.6 Multi-Agent Coordination

Purpose: Capture SA/MAP collaboration patterns.

Typical Sources:

• MAPSessionCompleted, MAPTurnCompleted events
• Session configuration and participant roles
• Handoff patterns

Example Use Case: Learn effective coordination strategies for multi-agent scenarios.

5. Compliance Boundaries

5.1 What is REQUIRED for v1.0 Compliance?

Aspect	Status	Notes
Schema definitions exist	REQUIRED	schemas/v2/learning/*.schema.json
Schemas are stable	REQUIRED	Frozen, backward-compatible
Samples conform to schemas (when emitted)	REQUIRED	Must validate against schemas

5.2 What is RECOMMENDED (Not Required)?

Aspect	Status	Notes
Runtime emits samples	RECOMMENDED	Not mandatory for v1.0
Collection at defined points	RECOMMENDED	Use standard collection points
Training/storage strategy	Out of scope	Product implementation decision

5.3 Summary

[!NOTE] Learning is OPTIONAL for v1.0 compliance. Runtimes MAY choose not to emit LearningSamples. However, IF samples are emitted, they MUST conform to MPLP schemas.

6. Implementation Guidance

6.1 Collection Strategy

1. Identify high-value collection points (see Collection Points)
2. Subscribe to trigger events via Event Bus
3. Extract relevant data from L2 objects and Observability events
4. Generate LearningSample conforming to schema
5. Store asynchronously (non-blocking)

6.2 Quality Labeling

Use the meta.human_feedback_label field for RLHF:

• approved: User accepted result
• rejected: User rejected result
• not_reviewed: No explicit feedback

6.3 Privacy

• MUST anonymize PII before storage if configured
• MUST support user opt-out
• SHOULD exclude sensitive files (e.g., .env, .secret)

7. Storage and Lifecycle (Out of Scope)

MPLP v1.0 does NOT specify:

• Where to store LearningSamples (filesystem, database, S3, etc.)
• How long to retain samples
• When to trigger training
• What models to train

These are product implementation decisions for TracePilot, Coregentis, or other runtimes.

8. Related Documents

Core Documentation:

• LearningSample Taxonomy - 6 families + triggers
• Collection Points - When to generate samples
• Learning Invariants - Validation rules
• Learning Sample Schema - Schema details
• Learning & Feedback - Detailed implementation patterns

Schemas:

• schemas/v2/learning/mplp-learning-sample-core.schema.json
• schemas/v2/learning/mplp-learning-sample-intent.schema.json
• schemas/v2/learning/mplp-learning-sample-delta.schema.json

Compliance:

• MPLP v1.0 Conformance Guide

---

Sample Families: 6 (intent, delta, pipeline, confirm, graph, coordination) Collection Points: 9 defined triggers Compliance: OPTIONAL for v1.0, schemas REQUIRED when emitting